[infurl]

https://www.zdnet.com/article/who-needs-a-supercomputer-your-desktop-pc-and-a-gpu-might-be-enough-to-solve-some-of-the-largest-problems/

Armed with a single turbocharged GPU, a team of researchers have successfully simulated part of a monkey's brain – something that normally requires a powerful and expensive supercomputer, but which the scientists maintain can now be carried out on a desktop PC.

The experiment, which involves simulating millions of neurons, as well as the billions of connections between those neurons, was performed by researchers from the University of Sussex using a single desktop PC and latest-generation Graphical Processing Units (GPUs).

If I understand the article correctly, they have developed a way to simulate the vast connectome of a brain by generating the connections on demand instead of storing all of them all the time. It's a bit like one of those games such as No Man's Sky that doesn't store a location until you have actually been there and caused it to be called into existence. It has been postulated that the entire universe could be a simulation that works just like that too.

[frankinstien]

I didn't see any mention of the type of storage of the states of the neuron(s) after it's finished processing. If it was a mechanical drive then what would be the improvement if they used an SSD? Also, can the approach be pre-emptive where you can have a cache of say 256GB for those areas that are about to be processed and swap out those cortical sections that are no longer needed into the SSD. This is what I do for my digital hippocampus which does employ that time chunking scheme that human brains use.

[infurl]

@frankinstien Previously, simulations on this scale required petabytes of RAM which is why they were limited to supercomputers. Even with resources like that, the article mentions that it took 8.4 minutes of computer time to simulate 1 second of brain time. It's possible that using SSD instead of RAM would not be that much slower since you still have the communication overheads between all those nodes anyway, but the main thing is that avoiding having to store the entire connectome by generating it algorithmically is much faster with far fewer resources.